Well maybe not that frequent (hundreds out of hundreds of thousands of packets) but I don't think I should see any in this (LAN) environment.
So I am doing a packet capture on my switch and I am seeing retransmissions to the Mellanox but I see the original packet in the capture. This leads me to believe that the Mellanox card is dropping the packet or at least not acking it. This happening on multiple boxes but not on all of our boxes (6 out of 10 that have ConnectX-3 EN cards). Some of the boxes with this issue are running VMA but I have observed it on one box that is not running VMA. The version of the driver doesn't seem to matter. I have seen it on boxes with: 2.2-1.0.0, 2.3-2.0.0, 2.4-1.0.4 and firmware 2.31.5050, 2.32.5100, 2.33.5000. I am also seeing it on servers running both Ubuntu 14.04 and servers running Centos 7. Neither the network nor the boxes are what I would consider overloaded.
Any ideas on what to look at?
Thanks. -- Bud